Merge pull request #334 from snowstormxd/codex/upstream-mkw-test

MKW performance improvements

1 files changed, 170 insertions, 0 deletions

diff --git a/src/tests/video_core/descriptor_table.cpp b/src/tests/video_core/descriptor_table.cpp
new file mode 100644
index 0000000000..8b20564949
--- /dev/null
+++ b/src/tests/video_core/descriptor_table.cpp
@@ -0,0 +1,170 @@
+// SPDX-License-Identifier: GPL-3.0-or-later
+
+#include <array>
+#include <chrono>
+#include <cstring>
+#include <iostream>
+#include <limits>
+#include <random>
+#include <catch2/catch_test_macros.hpp>
+#include "video_core/texture_cache/descriptor_table.h"
+
+namespace {
+struct DescriptorMemory {
+    std::array<std::array<u64, 128>, 2> banks{};
+
+    void ReadBlockUnsafe(GPUVAddr address, void* output, size_t size) const {
+        const auto bank = (address - 0x1000) / 0x1000;
+        const auto offset = (address - 0x1000) % 0x1000;
+        std::memcpy(output, reinterpret_cast<const u8*>(banks[bank].data()) + offset, size);
+    }
+};
+} // namespace
+
+TEST_CASE("Descriptor invalidation matches eager refresh across mutations", "[descriptor_table]") {
+    DescriptorMemory memory;
+    VideoCommon::DescriptorTable<u64> table;
+    REQUIRE(table.Synchronize(0x1000, 127));
+    std::array<bool, 128> seen{};
+    std::array<u64, 128> previous{};
+    std::mt19937 random{0x5080};
+    unsigned bank{};
+    u32 limit = 127;
+    for (unsigned i = 0; i < 40000; ++i) {
+        switch (random() % 5) {
+        case 0:
+            table.Invalidate();
+            seen.fill(false);
+            break;
+        case 1:
+            memory.banks[random() % 2][random() % 128] = random();
+            break;
+        case 2: {
+            const auto next_bank = random() % 2;
+            const u32 next_limit = random() % 2 ? 63 : 127;
+            const bool changed = next_bank != bank || next_limit != limit;
+            REQUIRE(table.Synchronize(0x1000 + next_bank * 0x1000, next_limit) == changed);
+            if (changed) {
+                seen.fill(false);
+            }
+            bank = next_bank;
+            limit = next_limit;
+            break;
+        }
+        default: {
+            const u32 index = random() % (limit + 1);
+            const u64 expected = memory.banks[bank][index];
+            const bool changed = !seen[index] || previous[index] != expected;
+            const auto actual = table.Read(memory, index);
+            REQUIRE(actual.first == expected);
+            REQUIRE(actual.second == changed);
+            seen[index] = true;
+            previous[index] = expected;
+        }
+        }
+    }
+}
+
+TEST_CASE("Repeated image deletion cannot reuse stale descriptor view IDs", "[descriptor_table]") {
+    DescriptorMemory memory;
+    std::array<VideoCommon::DescriptorTable<u64>, 5> channels;
+    std::array<std::array<unsigned, 128>, 5> view_ids{};
+    unsigned live_generation = 1;
+    const auto visit = [&](size_t channel, u32 index) {
+        if (channels[channel].Read(memory, index).second) {
+            view_ids[channel][index] = live_generation;
+        }
+        return view_ids[channel][index];
+    };
+    for (size_t channel = 0; channel < channels.size(); ++channel) {
+        REQUIRE(channels[channel].Synchronize(0x1000, 127));
+        for (u32 index = 0; index < 128; ++index) {
+            REQUIRE(visit(channel, index) == live_generation);
+        }
+    }
+    for (unsigned generation = 0; generation < 32; ++generation) {
+        ++live_generation;
+        for (unsigned deletion = 0; deletion < 50; ++deletion) {
+            for (auto& table : channels) {
+                table.Invalidate();
+            }
+        }
+        for (size_t channel = 0; channel < channels.size(); ++channel) {
+            for (u32 index = 0; index < 128; ++index) {
+                REQUIRE(visit(channel, index) == live_generation);
+                REQUIRE_FALSE(channels[channel].Read(memory, index).second);
+            }
+        }
+    }
+}
+
+TEST_CASE("Descriptor limits clear exposed slots after shrink and growth", "[descriptor_table]") {
+    struct AddressMemory {
+        void ReadBlockUnsafe(GPUVAddr address, void* output, size_t size) const {
+            std::memcpy(output, &address, size);
+        }
+    } memory;
+    constexpr std::array<u32, 9> limits{0, 63, 64, 127, 128, 0x7ffff, 0x80000, 0x80001, 0x100000};
+    for (const u32 limit : limits) {
+        VideoCommon::DescriptorTable<u64> table;
+        REQUIRE(table.Synchronize(0x1000, limit));
+        CHECK(table.Read(memory, limit).second);
+        CHECK_FALSE(table.Read(memory, limit).second);
+        table.Refresh(0x1000, 0);
+        table.Invalidate();
+        CHECK(table.Read(memory, 0).second);
+        CHECK_FALSE(table.Read(memory, 0).second);
+        table.Refresh(0x1000, limit);
+        CHECK(table.Read(memory, limit).second);
+        CHECK_FALSE(table.Read(memory, limit).second);
+        table.Invalidate();
+        table.Invalidate();
+        CHECK(table.Read(memory, limit).second);
+        CHECK_FALSE(table.Read(memory, limit).second);
+    }
+    for (const u32 limit :
+         {std::numeric_limits<u32>::max(), std::numeric_limits<u32>::max() - 0x1000}) {
+        VideoCommon::DescriptorTable<u64> table;
+        table.Refresh(0x1000, limit);
+        CHECK(table.Read(memory, 0).second);
+        table.Invalidate();
+        CHECK(table.Read(memory, 0).second);
+        CHECK_FALSE(table.Read(memory, 0).second);
+    }
+}
+
+TEST_CASE("Descriptor invalidation CPU benchmark", "[.][cache_bench]") {
+    DescriptorMemory memory;
+    const auto measure = [&](bool deferred) {
+        std::array<VideoCommon::DescriptorTable<u64>, 5> tables;
+        for (auto& table : tables) {
+            REQUIRE(table.Synchronize(0x1000, 127));
+        }
+        void (*volatile invalidate)(VideoCommon::DescriptorTable<u64>&) =
+            deferred ? +[](VideoCommon::DescriptorTable<u64>& table) { table.Invalidate(); }
+                     : +[](VideoCommon::DescriptorTable<u64>& table) {
+                           std::ranges::fill(table.read_descriptors, 0);
+                       };
+        size_t changed{};
+        const auto start = std::chrono::steady_clock::now();
+        for (unsigned batch = 0; batch < 2000; ++batch) {
+            for (unsigned deletion = 0; deletion < 50; ++deletion) {
+                for (auto& table : tables) {
+                    invalidate(table);
+                }
+            }
+            for (auto& table : tables) {
+                changed += table.Read(memory, 0).second;
+            }
+        }
+        const auto ms =
+            std::chrono::duration<double, std::milli>(std::chrono::steady_clock::now() - start)
+                .count();
+        REQUIRE(changed == 10000);
+        return ms;
+    };
+    const double eager = measure(false);
+    const double deferred = measure(true);
+    std::cout << "Descriptor bitmap invalidation: eager=" << eager << " ms deferred=" << deferred
+              << " ms (CPU only, not gameplay FPS)\n";
+}